Redaction cost the gate this shape on android: the driver typed the value twice
onto existing content, so the whole value is not its own doubling and the typed
value is not in the trace to compare against. The recorded-value check still
catches it on the backends that record one.
Red at this commit: G4 reports PASS on a field that grew by one string twice.
The setup skill's empty-page claim was the stale one that mattered: that run
records no_action_produced on every step and exits 1, it does not sit at
exit 0 with no violations. Numbers remeasured against the counter and
throwing fixtures.
The --paired flag advertised the signed-rank, two comments and a test message
still said rank-sum, and nothing said what rankSum is doing in the tree now
that no campaign reaches it.
The paired path had the same defect as the unpaired one: it subtracted two
step counts and handed the differences to the signed-rank test, so a pair
holding a run the wall clock stopped at step 12 entered as a difference
neither run supports. Twenty seeds where the first arm was still clean at
step 12 and the second violated at step 5 in six of them read sign -1 and
p 0.0011, pointing at the arm that never violated.
A pair is now scored the way the unpaired comparison scores one and tested by
the exact sign test over the pairs whose order censoring determines, which is
what the log-rank stratified by seed reduces to here. The signed-rank goes
with the differences it needed: a magnitude-based paired test wants a
difference from every pair, and the arms censor on different clocks. The
median difference stays, over the pairs where both runs violated, and says so.
a multi-key object selector concatenates its parts into one compound,
and a type selector is valid only at the head of one, so {id, secure}
built '[id="pwd"]input[type="password"]' and querySelectorAll threw.
Drives the built binary against a page with nothing to tap and reads the
process status, then the same run through campaign to pin what lands in
runs.jsonl: exit_code 1 there is a detection the analysis drops as missing
data.
jq splits an empty string into no segments, so reading the last one off an
action typed at coordinates threw and took the rest of the run's steps with it.
Such a step names nothing to check; the gate now passes over it and keeps
checking the ones that do.
An InputText that names no field aborts the analyzer, so the gate reports the
whole run as failed and checks none of the steps after it. 129 of the 485
recorded traces hold such a step.
Red at this commit: jq stops on a null selector and the gate reports FAIL.
it is the fourth fact the dump and the web runtime derive independently,
and the one that decides whether a typed value is written into the
shared record. three-valued, so the fixture guard requires all three
states rather than both polarities.
secure is derived from the field type, not written by the markup, so
matching it as a raw attribute reached nothing: the key is accepted, no
unknown-key error fires, and find answered undefined on web for the
field it answers with on ios. false is every editable field that is not
a password entry, since an element that is no field reports null and
answers to neither value.
The typed value stopped reaching the trace on any target that reports no
secure fact for the field, which on android is every field, so the gate was
comparing the redaction placeholder against itself and passing whatever the
driver did. The observed value is not redacted, and a field holding one string
twice over is the doubling itself. A value that is a single character repeated
stays exempt: the corpus types "a" 4096 times and a pair of spaces, and neither
can be told apart from its own doubling.
The recorded-value check stays for the targets that do record it, where it also
catches a doubling appended to content the field already held.
stepTimes threw the censoring flag away and handed the rank-sum a plain
number per run, so a run the wall clock stopped at step 12 was ranked as one
that violated at step 12. That was defensible while every clean run sat at
the budget, the largest value any run could take, and it stopped being
defensible when a clean run started being censored where it stopped.
Twenty runs clean at step 12 against twenty violations at step 100 read a12
0.000 and p 4.683e-10 from the rank-sum, in the same report as a log-rank
reading p 1.0000. The pairwise comparison is now the Gehan test over the
observations themselves, and the report says how many run pairs censoring
left with no order between them, which is how much of the effect size is the
null value rather than an observation.
The rank-sum carried over to right-censored samples: every pair of runs is
scored by which one outlived the other, and a pair censoring cannot order
counts as half rather than as a difference neither run supports. The effect
size and the p-value are the same statistic, and with nothing censored both
are exactly what the rank-sum reports.
The log-rank is one member of a family that differs only in how much each
event time counts. Nothing else changes: the counts it reports stay counts
whatever the weight, and the published-dataset results are unmoved.
--allow-no-properties was waiving two unrelated refusals, so a sweep passing
it for the property-free reason silently lost a detector it never asked to
disable, and a run with properties could only get the dead-run exemption by
claiming one it did not want.
A campaign never passes --exit-on-violation, so the refusal was discarding
runs that had found something: exit_code 1 in the record and the analysis
drops them as missing data. A run that recorded a violation holds a verdict,
which is the whole reason the refusal exists.
Android reports no secure fact for any field, so every InputText it records
writes the redaction placeholder rather than the typed value. The fixture still
carried the real value, which is the only reason the gate reported itself as
catching the doubling. Two more fixtures come with it: a repeated-character
corpus value that reads as its own doubling and must not fail, and a backend
that does record the typed value.
Red at this commit: G4 reports PASS on a doubled field it cannot see.
Gradle's install task talks to adb through ddmlib, which reads only
ANDROID_ADB_SERVER_PORT and dials the loopback address, so
ADB_SERVER_SOCKET never reaches it and `just test` could not touch a
remote emulator. Gradle now only assembles the APK and adb does the
install, which picks up the same server every other call in the run
talks to.
One arm's actions may include the login the spec's setup drove and the
other's cannot, so a per-action rate over the two divides by different
things and the tests rank the bookkeeping.
The full action key read the typed value straight from the trace, where
redaction renders every value typed into one field as the same string, so
two runs that typed different values there collapsed into one identity and
the report said nothing about it. The key now drops a redacted value, falls
back to the selector for that action, and counts the rows it did that to, so
the undercount reads as an undercount.
The setup exclusion landed for the model arm only, because only a model
pick stamped a source. A seeded run returned setup's action through the
same entry with no marker, so its denominator still counted the login
while the model arm's did not, and the two are compared.
serializeAction names setup and seeded on the wire, so both arms are
counted by one rule. An already-recorded trace names nothing and keeps
exactly the count it was reported with; unattributed_actions counts those
steps so the old denominator cannot pass as the new one. TraceVersion is
deliberately unbumped: oracle-reduction refuses a differing version, and
a bump would make all 169 recorded runs unreplayable.
folio extracts lastAction, and extractor values are persisted as
extractor_changes, so the password still reached the run directory
through the spec after the three render sites were closed.
The wrap sits in the runner rather than in lastActionFields because the
hosts hold the next step's tree, not the one the action was chosen
against: a field that stops being secure between the two would publish
what the trace withheld. Live and replay now agree byte for byte.
A folio login run wrote the account email and password in cleartext into
llm-calls.jsonl, 166 times in one run, beside screenshots of the same
screens. Three sites rendered it: the recent-action memory, the candidate
list, and the trace. One helper now covers all three so a fourth cannot
bypass it, and the driver still receives the real text.
Android redacts every typed value because it cannot tell a secure field
from any other. That asymmetry is deliberate and documented: safe by
default on the target that cannot tell.
ios reads it off SecureTextField, which the companion already sent and
nothing read; web reads input[type=password]. Android cannot: the native
tree mapper drops the password attribute before the sidecar sees it, so
the fact is three-valued and null there rather than a false that would
read as "not secure".
Defects per thousand actions divided by every dispatched step, so a
spec whose setup logs in inflated the denominator by however many steps
that took. It is the same error the run gate had, and it does not cancel
between arms.
A model run is separable because only an llm-selected action stamps
next_action.source. A seeded run is not: its setup returns through the
same entry with no marker, and 11261 dispatched steps across the 169
recorded runs carry no source at all, so excluding on it blind would
report every seeded run as having explored nothing. The seeded arm
counts as before and a test pins that.
Device was read from --device, which only an android run sets, so every
ios meta.json left the field empty and the trace could not say what
hardware produced it.
A green llm run carried no evidence of how much the generator actually
drove: the count was inferable only from llm-calls.jsonl outcomes, and
the number the refusal turns on was invisible in the run's own output.
A dead provider against folio exited 0 on a real emulator: the login
setup dispatched three actions before the generator was consulted, so
DispatchedActions was 3 and the gate never fired while the generator
drove the app zero times across 83 steps. Any spec with a login setup
was immune, which is the normal case.
Summary counts generator actions separately and the refusal reads that.
NoActionsDispatchedError becomes NoGeneratorActionsError, because a run
that dispatched three login taps was lying in the old name.
The run path grew the opt-out and the freeze gate did not, so a spec the
extraction and portability sweeps register nothing for on purpose could
be run but never frozen. The refusal now names the flag the way the
runner's does.
The line said the key matches placeholder alone, which is true of the web
runtime and not of the tree, where it resolves against the derived
attribute. A spec author reading it wrote a selector that matched on one
host and not the other.
Same argument as the zero-property refusal: an instrument that drove
nothing must not report a clean result. A first-screen violation still
wins under --exit-on-violation, --allow-no-properties exempts the
extraction sweeps that measure reach rather than judge, and one
dispatched action is enough, so a generator quiet on some screens is
untouched.
NextAction returning ErrNoAction left the step with no skip reason, so a
run whose every model call failed on transport, a non-2xx, an empty
choices array or an echo mismatch printed no violations and exited 0.
Only llm-calls.jsonl knew it had never touched the app.
The reason now travels the path the other five already take, so it
reaches the trace, the summary, and the campaign's dispatched-action
exclusion. A held step never asks and keeps carrying nothing.
A run stops at whichever comes first, the step budget or --duration, so
a clean run that reached the wall clock exited with fewer steps than the
budget and was still credited with the whole of it. The model arm pays a
network call and a screenshot per step, so it reaches the wall sooner and
was handed exposure it never had.
Nothing checked that two arms shared a budget either. Thirty identical
clean runs under budgets of 400 and 100 read a12 0.000 and p 1.685e-14
from the rank-sum while the log-rank in the same report read p 1.0000.
groupArms already refused this within one arm.
The claims the old convention left in comments and report lines are
corrected rather than left standing beside the new behaviour.
The sweep-level loop excluded a run on launch_error alone, while
excludedBecause already checked the campaign process's exit code. An
interrupted campaign wrote exit_code -1 with an empty launch_error, so
its one completed seed scored the implementation as a clean cell on a
tenth of the planned evidence.
The fixture builder wrote one exit code into both the sweep record and
the campaign run record, which is why no test could tell the two levels
apart.
The cache restored and the build ran anyway: a restored tarball keeps the
mtime it was archived with while checkout stamps the sources, so make read
every bundle as stale. Both logged Cache hit and rebuilt regardless.
Dating the bundles after their sources fixes the lie where it is told.
Order-only prerequisites would have fixed it in make, but a laptop has no
cache key, so editing prepare.sh would silently embed the previous tarball.
The formula version joins the key because a hit now decides what gets
embedded, and the key was blind to the brew install: the 1.1.8 and 1.5.0.b2
runs shared a key.
meta.json carried the host but not the device, so a trace could not say what
hardware produced it without the campaign manifest beside it. An experiment
splitting cells across api levels could only join them through that manifest.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
Preflight cannot catch a device that disappears mid-sweep, which is what
happened: the serials were alive the previous day. Three consecutive failures
under two minutes, with no run that worked in between, is a property of the
device and not a coincidence.
The manifest records which device was quarantined and which seeds have no
result, so an aborted sweep says so in its own artefact.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
A sweep launched at six serials, three of which had been deleted from the
host. 19 of 20 runs were lost, and not because half the devices were wrong:
a worker on a dead serial fails in about 31 seconds and immediately pulls
another seed, so three bad workers drained sixteen seeds while the three good
workers were still inside their first run.
Fast failure is more dangerous than slow failure, because the fast failure
consumes the resource the slow one would have left alone.
Preflight names every missing serial before the first seed is dispatched.
Claude-Session: https://claude.ai/code/session_01A5KmftdEJ49A9z5mF5ESrX
pj